Using the Bayesian credible subgroups method to identify populations benefiting from treatment: An application to the Look AHEAD trial.

Abstract

Traditionally, subgroup analyses are used to assess whether patient characteristics moderate treatment effectiveness with general disregard for issues of multiplicity. Using data from The Action for Health in Diabetes (Look AHEAD) trial in the United States, we aim to identify a subgroup where all of its types of members experience a treatment benefit defined as reducing the likelihood of a major cardiovascular event under an intensive lifestyle and weight-loss intervention. We apply the credible subgroups method to a Bayesian logistic model with a conservative prior that is sceptical of large treatment effect heterogeneity. The covariate profiles for which there is sufficient evidence of treatment benefit are, coarsely, middle-aged women, in poor subjective general health and with moderately to poorly controlled diabetes. There is at least 80% posterior probability that the conditional average treatment effect is positive for all covariate profiles fitting this description, which account for 0.5% of trial participants. Conversely, the covariate profiles that are likely to be associated with no benefit are middle aged and older men in excellent subjective general health, with well-controlled diabetes. These profiles apply to less than 2% of trial participants. More information is required to determine treatment benefit or no benefit for the remainder of the trial population.